Incremental class learning using variational autoencoders with similarity learning

Abstract Catastrophic forgetting in neural networks during incremental learning remains a challenging problem. Previous research investigated catastrophic fully connected networks, with some earlier work exploring activation functions and algorithms. Applications of have been extended to include similarity learning. Understanding how loss would be affected by is significant interest. Our investigates for four well-known similarity-based class The are Angular, Contrastive, Center, Triplet loss. results show that the rate differs across on multiple datasets. Angular was least affected, followed loss, Center good mining techniques. We implemented three existing techniques, iCaRL, EWC, EBLL. further proposed novel technique using Variational Autoencoders (VAEs) generate representation as exemplars passed through network’s intermediate layers. method outperformed state-of-the-art one does not require stored images (exemplars) generated representations from VAEs help preserve regions embedding space used prior knowledge so new “overwrite” it.